Developers can guide AI agents through verified scientific research procedures
Imagine telling an AI to predict a protein structure and having it find the necessary tools, execute them, and organize the results. This repository provides a collection of guidelines that help AI agents reduce common errors in scientific tasks and follow verified procedures. It is not tied to a specific programming language or framework and can be used with various AI tools that adhere to standards.
From Claude-specific to general-purpose scientific tools
The project recently changed its name to 'Scientific Agent Skills' and expanded beyond Claude support to work on any AI agent host that follows open standards, such as Cursor, Codex, and Google Antigravity. The latest version, 2.72.0, separates some document processing skills and updates the compatibility and verification procedures for the remaining 177 skills. Additionally, through K-Dense BYOK, a free open-source AI co-scientist that runs directly on the desktop, it provides a research workspace where user data remains local while accessing over 100 scientific databases and all skills.
177 verified research procedures
This repository includes 177 skills covering various fields such as biology, chemistry, medicine, physics, engineering, and earth sciences. Each skill specifies the scientific conventions and verification checklists required when using specific packages or data sources, guiding the AI to perform work that meets research standards rather than just writing code. For example, in bioinformatics, it provides specific guidelines for sequence analysis and single-cell RNA-seq processing, while in pharmaceuticals, it offers guidelines for molecular docking and pharmacokinetic modeling. These skills supply the context needed for AI agents to handle specialized scientific libraries and databases stably.
Database and tool accessibility
The core focus is helping AI agents find the information needed for research on their own. This repository provides a 'Database Lookup' skill that documents how to access more than 80 public or registered databases, including PubChem, ChEMBL, UniProt, and ClinicalTrials.gov. It also defines version-optimized workflows for over 70 Python packages, such as RDKit, Scanpy, and PyTorch Lightning, ensuring agents use the latest tools correctly. It also connects with nine scientific integration platforms, including Benchling and DNAnexus, allowing experiment data management and analysis pipelines to be executed with a single prompt.
Installation and environment setup
Installation can be performed simply using the npx skills add command or the GitHub CLI's gh skill command, depending on the supported host. It requires Python 3.13 or higher and the package manager uv, and works on macOS, Linux, and Windows (WSL2). Since each skill specifies its own dependencies and compatibility requirements, you should check the SKILL.md file for that skill after installation to install the necessary packages. It is recommended to selectively install only the skills needed for your research field rather than installing all skills at once.
Security review and reliability verification
Because AI agents can execute code and modify files, you must review the content of skills before installation. This repository regularly scans all skills using the Cisco AI Defense Skill Scanner and publishes the results, but there may be limitations in the review scope for community-contributed skills. Therefore, it is safer to check which external services and packages are used in SKILL.md and not install suspicious skills. In particular, skills that require network access may conflict with the user's system security policies, so it is recommended to go through a procedure to pre-approve access permissions for the relevant domains.
